Whether applications or libraries, today’s software heavily reuses existing code to build more gigantic software faster. To ensure a smooth user experience for an application’s end-user and a reliable software library for the developer, the shipped piece of software should be as bug-free as possible. Besides manual or automatic software testing, static program analysis is one possible way to find unintended behavior. While static analysis tools can detect simple problems using pattern matching, advanced problems often require complex interprocedural control- and data-flow analyses, which, in turn, presume call graphs. For example, call graphs enable static analyses to track inputs over method boundaries to find SQL-injections or null pointer dereferences. The research community proposed many different call-graph algorithms with different precision and scalability properties. However, the following three aspects are often neglected. First, a comprehensive understanding of unsoundness sources, their relevance, and the capabilities of existing call-graph algorithms in this respect is missing. These sources of unsoundness can originate from programming language features and core APIs that impact call-graph construction, e.g., reflection, but are not (entirely) modeled by the call-graph algorithm. Without understanding the sources of unsoundness’ relevance and the frequency in which they occur, it is impossible to estimate their immediate effect on either the call graph or the analysis relying on it. Second, most call-graph research examines how to build call graphs for applications, neglecting to investigate the peculiarities of building call graphs for libraries. However, the use of libraries is ubiquitous in software development. Consequently, disregarding call-graph construction for libraries is unfortunate for both library users and developers, as it is crucial to ensure that their library behaves as intended regardless of its usage. Third call-graph algorithms, are traditionally organized in an imperative monolithic style, i.e., one super-analysis computes the whole graph. Such a design can hardly hold up to the task, as different programs and analysis problems require the support for different subsets of language features and APIs. Moreover, configuring the algorithm to one’s needs is not easy. For instance, adding, removing, and exchanging support for individual features to trade-off the call graph’s precision, scalability, and soundiness. To address the first aspect, we propose a method and a corresponding toolchain for both a) understanding sources of unsoundness and b) improving the soundness of call graphs. We use our approach to assess multiple call-graph algorithms from state-of- the-art static analysis frameworks. Furthermore, we study how these features occur in real-world applications and the effort to improve a call graph’s soundness. Regarding aspect two, we show that the current practice of using call-graph algorithms designed for applications to analyze libraries leads to call graphs that both a) lack relevant call edges and b) contain unnecessary edges. Ergo, motivating the need for call-graph construction algorithms dedicated to libraries. Unlike algorithms for applications, call-graph construction algorithms for libraries must consider the goals of subsequent analyses. Concretely, we show that it is essential to distinguish between the analysis’s usage scenario. Whereas an analysis searching for potentially exploitable vulnerabilities must be conservative, an analysis for general software quality attributes, e.g., dead methods or unused fields, can safely apply optimizations. Since building one call graph that fits all needs is nonsensical, we propose two concrete algorithms, each addressing one use case. Concerning the third aspect, we devise a generic approach for collaborative static analysis featuring modular analysis that are independently compilable, exchangeable, and extensible. In particular, we decouple mutually dependent analyses, enabling their isolated development. This approach facilitates highly configurable call-graph algorithms, allowing pluggable precision, scalability, and soundiness by either switching analysis modules for features and APIs on/off, or exchanging their implementations. By addressing these three aspects, we advance the state-of-the-art in call-graph construction in multiple dimensions. First, our systematic assessment of unsoundness sources and call-graph algorithms reveals import limitations with state-of-the-art. All frameworks lack support for many features frequently found in-the-wild and produce vastly different CGs, rendering comparisons of call-graph-based static analyses infeasible. Furthermore, we leave both developers and users of call graphs with suggestions that improve the entire situation. Second, our discussion concerning library call graphs raises the awareness of considering the analysis scenario and opens up a new facet in call-graph research. Third, by featuring modular call-graph algorithms we ease to design, implement, and test them. Additionally, it allows project-based configurations, enabling puggable precision, scalability, and sound(i)ness.